The planning of adequately powered research designs increasingly goes beyond determining a suitable sample size. More challenging scenarios demand simultaneous tuning of multiple design parameter dimensions and can only be addressed using Monte Carlo simulation if no analytical approach is available. In addition, cost considerations, for example, in terms of monetary costs, are a relevant target for optimization. In this context, optimal design parameters can imply a desired level of power at minimum cost or maximum power at a cost threshold. We introduce a surrogate modeling framework based on machine learning predictions to solve these optimization tasks. In a simulation study, we demonstrate the efficiency for a wide range of hypothesis testing scenarios with single- and multidimensional design parameters, including t tests, analysis of variance, item response theory models, multilevel models, and multiple imputations. Our framework provides an algorithmic solution for optimizing study designs when no analytic power analysis is available, handling multiple design dimensions and cost considerations. Our implementation is publicly available in the R package mlpwr. (PsycInfo Database Record (c) 2023 APA, all rights reserved).

中文翻译：

---

基于仿真的统计功效设计优化：利用机器学习。


充分有力的研究设计的规划越来越超出了确定合适的样本量的范围。更具挑战性的场景需要同时调整多个设计参数维度，并且如果没有可用的分析方法，则只能使用蒙特卡罗模拟来解决。此外，成本考虑（例如货币成本）也是优化的相关目标。在这种情况下，最佳设计参数可以意味着以最小成本获得期望的功率水平或者以成本阈值获得最大功率。我们引入基于机器学习预测的代理建模框架来解决这些优化任务。在模拟研究中，我们展示了使用单维和多维设计参数进行各种假设检验场景的效率，包括 t 检验、方差分析、项目响应理论模型、多级模型和多重插补。我们的框架提供了一种算法解决方案，用于在没有分析功效分析可用时优化研究设计，处理多个设计维度和成本考虑。我们的实现在 R 包 mlpwr 中公开可用。 （PsycInfo 数据库记录 (c) 2023 APA，保留所有权利）。